Levandor

Weather Strategy Verdicts — US/EU/AS (2026-07-02)

6 min read

Post-adversarial lane verdicts from the 2026-07-02 weather-strategy synthesis (14-agent workflow: 6 recon scouts → 4-angle strategy panel → 1 skeptic per design), grounded in live Gamma/CLOB pulls, a fresh Jun 14–Jul 1 backtest across 19 cities, and 6-season lock-hazard climatology. Two headline results were overturned in verification — see weather-lookahead-trap-2026-07-02. Canonical detail: /Users/levander/coding/pmv2/research/weather-strategy-synthesis-2026-07-02.md.

For Agents

These verdicts supersede prior lane dispositions across us_weather_algo / as_weather_algo / eu_weather_algo. Prior beliefs overturned today are in weather-prior-corrections-2026-07-02. All PnL evidence is 15-min sampled quote prints — fill quality is the binding unknown for anything taker-shaped.

Verdict table

LaneVerdictOne-line reason
US taker lock-and-holdPILOT NOW ($100 clips, 4+ wks)Only region with fillable depth + intact 40–123 min repricing lag; clean causal backtest (n=67, +29.9%/trade) still stands
EU taker lock-and-holdDROP$8 median fillable-within-3c depth; whales sweep the liquid moment in 1 min
EU maker / LR farmingPILOT after infra (cancel() + GTD)madrid/paris quoting vacuums; $703–828/day city LR pools are real
Asia takerRESEARCH-MORE, no capitalHeadline +0.455/; one free replay decides CN/Tokyo
Cheap tails / adjacent-flip-tail buyingDEAD (empirical)1 winner / 3,310 tail-buckets; +1°C tail 0.9% win at 0.049 avg (~5× overpriced)

US taker lock-and-hold — PILOT NOW

  • Cities: nyc, atlanta, chicago, miami (denver excluded, reach@15 = 0.353; miami on probation at 0.579).
  • Signal: existing causal lock — ≥15:00 station-local, 90-min whole-°F no-rise; truth feed = WU poll + METAR-tenths fallback, no-obs >45 min → no entries.
  • Gate: locked bucket’s own market price ≥ 0.60 at lock. No ML gate — raw entry price (AUC 0.714) beats the trained GBT (0.684).
  • Entry: FAK taker, cap 0.95; pre-send book re-check, reject if ask > signal price + 2c.
  • Exit: hold to settlement. No stops (realistic stop−0.05 = −0.70/$ — gaps). No flatten overlay (test-set artifact). PT 0.92: A/B 50/50 or drop (≥0.90 in final 2–3h resolved YES 55/55 → PT mostly sells winners early).
  • Size: flat **1,500 — the only binding control for the correlated NYC/ATL/MIA gap-night.
  • Blocker before first trade: fee-accrual fix (fee_usd=0 in position_manager) — pilot scoring is invalid without it. See [[weather-prior-corrections-2026-07-02#4-fee_usd0-ledger-bug-overstates-live-roi|4. fee_usd=0 ledger bug overstates live ROI]].
  • Pass gate (n≥50 fills, ~6–9 wks): roi_lcb_80 > 0 net of true fees AND median slippage ≤2c AND wr ≥75% AND fill-vs-signal ≤2c on ≥80% of fills. Any FAIL → back to shadow, no re-tuning on pilot data.
  • Standing monitors: weekly realized repricing lag (retire lane if median <20 min — the 40–123-min lag is competitor inattention, not physics; HondaCivic laps at 4.7 min when attending); SPRT wr floor; settlement audit.
  • Honest EV: unknown-sign until fills measured; best estimate −2c to +8c/$ net. The pilot is a measurement, not an income claim.

EU taker lock-and-hold — DROP (do not build eu_weather_algo for this)

  • Fillable-within-3c depth (D3c) at the money: **6; madrid: 100–300 clip cannot execute; taking whole books yields <$50/day EU-wide.
  • The liquid moments belong to whales: london Jul-1 repriced 0.72→0.90 in 1 minute behind a $460 sweep. You’d be the adverse selection.
  • Best gated cohort: +0.083/$ at mid-print fantasy fills, n=17 (~14 independent city-days, one record-heat fortnight); Wilson-80 wr lower bound is below breakeven.
  • Heat-wave °C flips at conviction 0.84–0.91; at forced entries ≥0.80, one miss erases 5+ wins.

Reopen monitor (build this — ~1h, zero risk)

Daily EU book snapshots of modal+adjacent buckets at 16:00–20:30 local for 2 weeks. Reopen only if D3c ≥ $300 on ≥50% of days AND first sweeper arrival reliably >15 min post-lock. (Today’s depth snapshot was taken 2–5.5h pre-window, n=1 day — the competition plank carries the DROP.)

EU maker / LR farming — PILOT after infra (the actual EU opportunity)

  • madrid + paris are quoting vacuums (12–18c spreads); city LR pools $703–828/day + 25% maker rebate — see weather-prior-corrections-2026-07-02.
  • Blocked on position_manager builds: GTD/GTC param in build_order; cancel() live-test (50 reps place+cancel deep GTD, p95 <5s, 0 orphans); maker quote manager + self-cross guard (wash-trade ToS).
  • Spec (skeptic-corrected): 10 days, 2 markets, inventory cap $150–200, two-sided ≥50 shares inside 4.5c, no modal-bucket quoting during the 15:00–18:00 flip window (LFPB 15:00 hazard .613 — nowcaster flow picks you off); quote late-evening/overnight and far buckets.
  • Pass: LR accrual ≥ 2× inventory losses AND cancel p95 <5s AND no day-loss >$150 AND zero self-crosses. Pool-split formula unverified — measure, don’t assume.

Asia — RESEARCH-MORE (no capital, no wiring)

  • Seoul: fold. 65% of days locked out (book >0.95 before any entry could exist); repriced in 1 min on Jul-1 before a lock existed; no LR pool on highest. Revival only via KMA APIhub 1-min AMOS feed beating the 30-min-METAR watchers.
  • CN + Tokyo: one binding free experiment decides. Lock-timed replay — IEM METARs for RJTT/ZHHH/ZBAA/ZGSZ, actual ≥15:00/90-min-no-rise lock times on hourly-METAR semantics, enter at first curve price after lock, cap 0.92, gate at lock time; PLUS one day of CN/RJTT book snapshots 06–10 UTC (CN entry-zone depth has never been measured). Pilot bar: mean ≥ +0.10/100.
  • Causal proxy context: +0.145/ at engine-feasible ≥16:30** (n=10) — the known “later intl lock” trap.
  • Tokyo asset: JMA AMeDAS 44166 (10-min, 0.1°C, no key, ~3–10-min lag) — the speed city if the replay passes; but tokyo lockout is 44%, nearer Seoul than Wuhan.
  • Timezone-ladder redemption chain (AS→US→EU) is real but pointless until ≥2 lanes have positive expectancy.

Confirmed DEAD (do not re-explore)

  • Cheap late tails (≤0.13 at 20:00 local): 1 winner / 3,310, sum −1.64/$.
  • Adjacent +1°C bucket as flip insurance: n=117, 0.9% win at 0.049 — that tail’s overpricing IS the lock trade’s edge.
  • Also dead: reactive price stops; PT ≥0.95; trained P(win) models vs raw entry price; hazard-gating early EU locks (tenths gate recovers only +1–1.8pp — lock HOUR is the lever); Hong Kong (0.1°C HKO); Istanbul/Moscow (no public tenths); convergence-conditioned hit rates (1.96× inflation); the C1/C2/C5/C6/C8/C10–C14 graveyard from edge-discovery-2026-06-22.

Portfolio + capacity

150–200 inventory once unblocked; AS 0%; rest settlement float. Taker capacity ceiling ≈ **15k cannot deploy in this family as a taker.

Build list (position_manager, priority order)

  1. Fee accrual (fee_usd=0 at lib.rs ~L521) — blocks US pilot scoring
  2. PT trigger → existing ExitIntent plumbing — only if PT survives the A/B
  3. GTD/GTC param in build_order — blocks maker
  4. cancel() live-test (dead code at lib.rs ~L531) — blocks maker
  5. Region-day cap + 2nd-flip region breaker in reserve.rs
  6. Obs-staleness kill-switch
  7. v2 envelope NATS leg (AS lane’s leg was never wired; EU/US converge on v2)
  8. AS bug fixes M4/M5/M6/L2/H5 before any AS revival
  9. Maker quote manager + self-cross guard

Caveats

All market-price evidence is May 5 – Jul 1 2026 (one early-summer regime); July–Aug convection (ATL/MIA) and mei-yu/typhoons (CN) untested. Every PnL number is from 15-min sampled prints, never real asks. Competitor state is a snapshot — one config change on their side ends the US lane, hence the standing lag monitor.

Graph View

Backlinks